Re: VQ and Huffman codebooks creation

On Sun, Aug 16, 2009 at 3:58 PM, Sampo Syreeni<[email protected]> wrote:
> On 2009-08-16, Ralph Giles wrote:
>
>> In the libvorbisenc implementation the codebooks are pre-generated
>> based on training data, and a set are selected based on the quality
>> setting.
>
> Or in other words, while Vorbis allows variable codebooks, currently the
> codebook is constant over all of the encoded files. It's an average case
> codebook, statically compiled against a typical training set of tracks,
> not something that is adapted separately to each and every encoded
> track. Adapting the codebook on the fly would be something of a research
> challenge.

Adjusting the huffman tables on the fly, OTOH, would be pretty
straight forward. But there are downsides... like the inability to
simply convert chained streams into non-chained ones. The savings is
pretty small too.